On Thu, 12 Mar 2026 13:16:30 GMT, Benoît Maillard <[email protected]> wrote:
> This PR fixes an IR mismatch in `TestLWorld.java` for `test85` and `test86`. > When merging jdk-26+17, there were some changes to > `G1BarrierSetC2::estimated_barrier_size` due to > [JDK-8342382](https://bugs.openjdk.org/browse/JDK-8342382). The barrier size > is significantly reduced, opening the door to more unrolling as per > `IdealLoopTree::policy_unroll`. In the specific case of `test85` and > `test86`, this leads to unrolling where there previously was none and > splitting a `CountedLoop` into a main and post loop. This breaks the IR rules > because they rely on exact nodes count. > > For `test86`, we can trivially replace `COUNTED_LOOP` by `COUNTED_LOOP_MAIN` > to avoid counting the loop twice. > > For `test85`, we also have a check on `LOAD_UNKNOWN_INLINE`, which ends up > being appearing four times in the default setup (because of unrolling and > pre/main/post). This is a bit annoying. It seems the best solution in this > case is to check the IR right before the pre/main/post, that is at > `PHASEIDEALLOOP2`. > > ### Testing > - [x] GitHub Actions > - [x] tier1-4 for this specific test, plus some internal testing > > Thank you for reviewing! @benoitmaillard Your change (at version 65de21d9705776a9d24a1fee51091a074dd73a64) is now ready to be sponsored by a Committer. ------------- PR Comment: https://git.openjdk.org/valhalla/pull/2219#issuecomment-4065675256
